Research map / Adversarial robustness and security / Attacks on text and graphs
DIFFender: Diffusion-Based Adversarial Defense against Patch Attacks
Caixin Kang, Yinpeng Dong, Zhengyi Wang, Shouwei Ruan, Yubo Chen, Hang Su, Xingxing Wei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-entropy mask loss L_CE to prompt tuning (ablation) | DIFFender · ImageNet BPDA+AdvP | robust accuracy (%): 76.2 -> 88.3 | Improved |
| add L1 restoration loss to prompt tuning (ablation) | DIFFender · ImageNet clean | accuracy (%): 88.3 -> 91.4 | Improved |
| add feature-distance loss d to prompt tuning (ablation) | DIFFender · ImageNet BPDA+AdvP | robust accuracy (%): 87.1 -> 88.3 | Improved |
| add patch restoration after localization (vs. zeroing masked region) (ablation) | DIFFender · ImageNet BPDA+AdvP | robust accuracy (%): 84.0 -> 88.3 | Improved |
| use few-shot tuned prompts instead of empty prompt (ablation) | DIFFender · ImageNet BPDA+AdvP | robust accuracy (%): 76.4 -> 88.3 | Improved |
| use few-shot tuned prompts instead of manual prompts (ablation) | DIFFender · ImageNet BPDA+AdvP | robust accuracy (%): 77.9 -> 88.3 | Improved |
| use plain DiffPure purification against patch attacks | Inception-v3 · ImageNet BPDA+AdvP | robust accuracy (%): 10.5 | Hurt |
| tune prompts on Inception-v3 only, transfer to other classifiers | DIFFender · ImageNet ResNet-50 AdvP | robust accuracy (%): Jedi 46.9 -> DIFFender 83.2 | Improved |
| use DIFFender vs Jedi/SAC on larger patches | DIFFender · ImageNet AdvP 15% patch | robust accuracy (%): SAC 34.8 -> DIFFender 56.6 | Improved |
| apply feature-norm clipping defense to a transformer classifier | FNC · ImageNet BPDA+AdvP | robust accuracy (%): Inception-v3 61.3 -> Swin-S 6.3 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Jedi — Jedi: Entropy-Based Localization and Removal of Adversarial Patches (CVPR 2023) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.89)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.88)
- PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies (CVPR 2025 · similarity 0.88)
- DiffAdvMAP: Flexible Diffusion-Based Framework for Generating Natural Unrestricted Adversarial Examples (ICML 2025 · similarity 0.87)
- AdvAD: Exploring Non-Parametric Diffusion for Imperceptible Adversarial Attacks (NeurIPS 2024 · similarity 0.86)
- CausalDiff: Causality-Inspired Disentanglement via Diffusion Model for Adversarial Defense (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
